Water, wind and light storage multi-target scheduling method based on improved hybrid particle swarm grey wolf optimization algorithm

Abstract

The invention belongs to the technical field of power system optimization scheduling, and particularly relates to a multi-target day-ahead peak regulation scheduling method for a cascade water-wind-solar-storage complementary power generation system. The method comprises the following specific steps: establishing a mathematical model of the water-wind-solar-energy-storage complementary system, and constructing a multi-target optimization framework considering power generation benefits and peak regulation effects; constructing a prediction network by using an adaptive multi-time scale Improved LSTNet to carry out day-ahead prediction on the incoming water amount of the cascade basin interval, the generated power of the photovoltaic power station and the generated power of the wind power plant, and taking the predicted data as input data of the model; an improved multi-objective hybrid particle swarm grey wolf optimization algorithm (IMOPSO-GWO) is adopted to solve the established model, global exploration and local development capabilities are coordinated through a convergence factor adaptive adjustment mechanism, a dynamic mutation operator is introduced to enhance Pareto leading edge search efficiency, and finally a day-ahead scheduling scheme of the complementary system is obtained. By optimizing the output of the cascade hydroelectric generating set and the energy storage, the method has obvious effects on reducing the residual load fluctuation of the power grid and improving the power generation income.

Description

Technical Field

[0001] The present invention belongs to the technical field of optimal dispatching of power systems, and particularly relates to a multi-objective day-ahead peak shaving dispatching method for a cascade hydropower-wind-solar-storage complementary power generation system. Background Art

[0002] With the accelerated construction of a new power system, the large-scale grid connection of renewable energy with water energy, wind energy, and solar energy as the core has become a key path for the transformation of the energy structure. The cascade hydropower-wind-solar-storage integrated complementary system can effectively suppress the output fluctuations of new energy, improve the peak shaving capacity and power supply reliability through multi-energy collaborative optimization, and is an important carrier for realizing the integrated operation of the power source, grid, load, and storage. As of 2022, as the largest clean energy base in the country, the hydropower installed capacity in Sichuan Province accounts for more than 75%, and the new energy installed capacity accounts for 6.1%. Conducting research on the optimal dispatching of cascade hydropower-wind-solar-storage can not only improve the utilization efficiency of water resources but also give full play to the regulating role of reservoirs and maximize economic benefits.

[0003] The short-term operation of cascade hydropower stations is a complex multi-stage optimization problem with multi-dimensional, multi-constrained, non-linear, and dynamic characteristics. The main task is to comprehensively consider the reservoir water level, inflow, and the actual operation of the power grid in the electricity market environment, and study the optimal operation and load distribution of hydropower stations to ensure the safe and economic operation of hydropower stations. On the premise of ensuring the stable operation of the power grid, the utilization efficiency of water energy resources is maximized through refined load distribution. CN118801486B proposes a short-term risk dispatching method for hydropower-wind-solar that takes into account the utilization of water resources in the remaining period, and CN116667393A proposes a collaborative optimization method for the day-ahead power generation plan and flexibility response rules of cascade hydropower-wind-solar, providing a new technical path for short-term operation optimization.

[0004] With the advancement of the electricity market reform, power plants and power grids are separated for operation. As the main power generation entities, cascade hydropower plants need to make profits and losses on their own and pursue the maximization of power generation benefits. At the same time, they need to undertake peak shaving and frequency modulation tasks to ensure the power supply safety of the power grid. This game relationship between economy and security has led to a sharp increase in the difficulty of formulating multi-time scale operation plans. That is, excessive pursuit of power generation benefits will weaken the peak shaving capacity of the power grid, while emphasizing peak shaving tasks is likely to cause waste of water energy resources. There is an urgent need to construct an optimization decision-making mechanism that takes into account the coordination between the grid and the power source. In this context, the development of collaborative dispatching technology for cascade hydropower-wind-solar-storage systems that integrates multi-objective optimization and intelligent decision-making has become the core proposition for breaking through the bottleneck of efficient utilization of water resources and ensuring the safe and economic operation of the new power system, with both theoretical research value and engineering practice significance. Summary of the Invention

[0005] The objective of the present invention is to provide a multi-objective day-ahead peak shaving scheduling method for a cascade hydropower, wind, and solar energy storage complementary system based on an improved hybrid particle swarm grey wolf optimization algorithm, to solve the multi-objective collaborative optimization problem in the scheduling of cascade hydropower stations and new energy, improve the peak shaving capacity and economic benefits of the system, and ensure the safe and stable operation of the power grid.

[0006] The technical solution adopted by the present invention is as follows: A multi-objective day-ahead peak shaving scheduling method for a cascade hydropower, wind, and solar energy storage complementary system based on an improved hybrid particle swarm grey wolf optimization algorithm is specifically implemented according to the following steps:

[0007] Step 1: Establish a mathematical model of the integrated hydropower, wind, and solar energy storage complementary system, comprehensively consider the constraints of UHV DC tie lines, cascade hydropower constraints, energy storage constraints, and demand response constraints, and construct a multi-objective optimization framework for cascade hydropower, wind, and solar energy storage considering power generation benefits and peak shaving effects.

[0008] Step 2: Propose a day-ahead prediction framework for hydropower, wind, and solar energy based on an adaptive multi-time scale Improved LSTNet. Use multi-scale convolutional kernels to design and extract features of different scales in the time series, enhancing the model's ability to capture multi-level time series patterns such as different periodicities, trends, and randomness. At the same time, combine an adaptive convolution mechanism to make dynamic adjustments according to data features.

[0009] Step 3: Propose an improved multi-objective hybrid particle swarm grey wolf optimization algorithm (IMOPSO-GWO) to solve the established model. Use Tent-Logistic-Bernoulli multiple chaotic mappings to initialize the grey wolf population, coordinate the global exploration and local exploitation capabilities through a convergence factor adaptive adjustment mechanism, and introduce a dynamic Gaussian mutation operator to enhance the Pareto front search efficiency, finally obtaining a multi-objective day-ahead scheduling plan for the complementary system.

[0010] The characteristics of the present invention also lie in:

[0011] Step 1 is specifically implemented as follows:

[0012] In peak shaving scheduling, the average value of the absolute value of the surplus load anomaly is an important indicator for evaluating the quality of peak shaving effects. It represents the fluctuation degree of the remaining load, that is, the deviation degree of the remaining load from the ideal smooth state. The smaller the average value of the absolute value of the surplus load anomaly, the smaller the fluctuation of the remaining load and the better the peak shaving effect; conversely, it indicates a poorer peak shaving effect. Therefore, in the short-term scheduling optimization of cascade hydropower stations, this indicator is selected as the objective function for power grid peak shaving.

[0013]

[0014] In the formula: F1 is the sum of the average values of the absolute values of the surplus load anomalies of all power grids; G is the number of power grids; T is the total number of scheduling periods; L is the number of UHV DC tie lines; Dg,t and D' g,t are the system load and the remaining load of the g - th power grid in period t respectively; P l,g,t is the power transmitted from the cascade hydro - wind - solar - storage integrated complementary system to the g - th power grid by the tie - line l in period t; is the demand response power of the g - th power grid in period t.

[0015] The optimal scheduling strategy with the maximum cascade hydropower generation as the objective function aims to maximize the power production benefit by reasonably allocating the power generation tasks of each hydropower station, making full use of the available water resources, and enabling the overall power generation capacity of the cascade reservoir system to be optimally exerted.

[0016]

[0017] In the formula: F2 is the total cascade hydropower generation; I is the number of cascade hydropower stations; is the output of the n - th unit of hydropower station i in period t; N i is the number of units of hydropower station i.

[0018] Constraints of the UHVDC tie - line for the multi - objective optimization model of the hydro - wind - solar - storage integrated complementary system:

[0019] (a) Upper and lower limits constraints of the tie - line power

[0020] P l,min ≤P l,t ≤P l,max (3)

[0021] In the formula: P l,t is the transmission power of the tie - line l in period t; P l,min and P l,max are the minimum and maximum values of the transmission power of the tie - line l respectively.

[0022] (b) Variation range constraints of the tie - line power

[0023] |P l,t -P l,t-1 |≤ΔP l (4)

[0024] In the formula: ΔP l is the maximum fluctuation amplitude of the transmission power of the tie - line l in adjacent periods.

[0025] Constraints of the cascade hydropower for the multi - objective optimization model of the hydro - wind - solar - storage integrated complementary system:

[0026] (a) Water balance constraint

[0027]

[0028] In the formula: Vi,t is the storage capacity of power station \(i\) at the end of time period \(t\); \(I\) i,t is the inflow of power station \(i\) in time period \(t\); \(\tau\) is the water flow lag time between power station \(i\) and its upstream power station \(i - 1\); \(Q\) i,t is the outflow of power station \(i\) in time period \(t\); \(Q\) i-1,t-τ is the flow of power station \(i - 1\) at time \(t-\tau\) considering the water flow lag time; \(R\) i,t is the sectional flow between power station \(i - 1\) and power station \(i\) in time period \(t\); and are the power generation flow and the water spillage flow of power station \(i\) in time period \(t\) respectively; is the power generation flow of unit \(n\) of power station \(i\) in time period \(t\); \(U\) i,n,t represents the start - stop status flag of the \(n\)th unit of hydropower station \(i\) in time period \(t\), which is a \(0 - 1\) variable.

[0029] (b) Reservoir water level limit constraint

[0030]

[0031] In the formula: is the water level of the reservoir of hydropower station \(i\) at the end of time period \(t\), and are the upper and lower limits of the water level of the reservoir of hydropower station \(i\) respectively.

[0032] (c) Outflow constraint

[0033] Q i,min ≤Q i,t ≤Q i,max (7)

[0034] In the formula: \(Q\) i,max and \(Q\) i,min are the upper and lower limits of the outflow of hydropower station \(i\) respectively.

[0035] (d) Power station output limit constraint

[0036]

[0037] In the formula: is the output of power station \(i\) in time period \(t\); and are the upper and lower limits of the output of power station \(i\) respectively.

[0038] (e) Water level - storage capacity relationship

[0039]

[0040] In the formula: is the non - linear relationship curve function of the water level - storage capacity of the reservoir where power station \(i\) is located.

[0041] (f) Tail water level - discharge relationship

[0042]

[0043] Where: is the non - linear relationship curve function between the tail water level and the discharge of power station i; is the tail water level of power station i at time period t.

[0044] (g) Generator head constraint of the unit

[0045]

[0046] Where: H i,n,t and are respectively the generator head and the head loss of the n - th unit of power station i at time period t.

[0047] (h) Head loss function

[0048]

[0049] Where: α i and β i are respectively the head loss coefficient and the loss constant of power station i, which can generally be obtained through hydraulic tests.

[0050] (i) Unit output characteristic relationship

[0051]

[0052] Where: η i is the unit efficiency of power station i; ρ is the density of water; g is the acceleration due to gravity;

[0053] (j) Unit output constraint

[0054]

[0055] Where: and are respectively the upper and lower limits of the output of the n - th unit of hydropower station i.

[0056] (k) Unit output ramp - up constraint

[0057]

[0058] α i,n,t +β i,n,t ≤1(16)

[0059] Where: is the maximum fluctuation amplitude between adjacent time periods of the n - th unit of power station i. α i,n,t ∈{0,1} and β i,n,t∈{0,1} is the power regulation index variable of the nth unit of power station i in period t, α i,n,t = 1 indicates that the power is regulated downward in period t + 1, β i,n,t = 1 indicates that the power is regulated upward in period t + 1. When the power does not change, α i,n,t = 0 and β i,n,t = 0.

[0060] Energy storage constraints of the multi-objective optimization model of the integrated water-wind-solar-storage complementary system:

[0061] (a) Energy storage power constraint

[0062]

[0063] In the formula: are the charging and discharging powers of the battery in period t respectively, are the maximum and minimum values of the battery charging and discharging powers respectively, are the charging and discharging state flag bits of the battery respectively. The value of 1 indicates operation, and the value of 0 indicates stop.

[0064] (b) Energy storage capacity constraint

[0065]

[0066] In the formula: is the remaining capacity of the battery at the end of period t, η ES,cha 、η ES,dis are the charging and discharging efficiencies of the battery respectively, are the maximum and minimum values of the remaining capacity of the battery respectively.

[0067] (c) Energy storage charge and discharge frequency constraint

[0068]

[0069] In the formula: N is the maximum number of charge and discharge times of the battery within a scheduling period.

[0070] Demand response constraints of the multi-objective optimization model of the integrated water-wind-solar-storage complementary system:

[0071]

[0072] In the formula: γ is the maximum adjustable power coefficient of the single-period demand response; is the total maximum adjustable power of the demand response within the scheduling period.

[0073] Step 2 is specifically implemented as follows:

[0074] Collect relevant historical data and meteorological data for cascade hydropower, wind power, and photovoltaic power prediction, including: historical hydrological data of the basin area, power generation power of photovoltaic power plants and wind farms, rainfall, light intensity, wind speed, temperature, etc., and organize the above data into the same time window. Subsequently, outlier removal and missing value filling are performed. In the present invention, the Z-score method is used to perform outlier removal processing on the collected data, and when the relative difference is greater than 3, the data is removed as an outlier.

[0075]

[0076] Where: X is the sample data, μ is the sample mean, and σ is the standard deviation.

[0077] For blank values, inverse distance weighted interpolation is used and combined with the reasonable guidance range of meteorological data and expert experience screening for filling. Since the above time series data has significant seasonal fluctuations, masking the underlying trends or periodic changes. Therefore, through STL seasonal decomposition (Seasonal-Trend Decomposition using LOESS), the seasonal fluctuations are separated from the original data, enabling the model to focus more on the data trends and residual parts, thereby improving the prediction accuracy.

[0078] The core of the STL method lies in decomposing the time series Y t into three parts. The trend component is estimated by smoothing the original data through Loess regression; the seasonal component is obtained by Loess smoothing after removing the trend component; the residual component is the result of subtracting the trend and seasonal components from the original data.

[0079] Y t = T t + S t + R t (22)

[0080] T t = Loess(Y t ) (23)

[0081] S t = Loess(Y t - T t ) (24)

[0082] R t = Y t - T t - S t (25)

[0083] Where: Y t is the original time series data; T t is the trend component; S tis the seasonal component; R t is the residual component.

[0084] To make the training of the Improved LSTNet model more stable, the Min-Max normalization method is applied to normalize the input data.

[0085] Using the multi-scale convolutional kernel design, features of different scales of the time series are extracted to enhance the model's ability to capture multi-level time series patterns such as different periodicities, trends, and randomness. Set the time series data X = {x1, x2, … x T}, where is the input vector at time t, d is the input feature dimension, and T is the data length. The LSTNet model simultaneously applies convolutional kernels K1, K2, …, K m of different sizes, and the convolutional operations are performed in parallel.

[0086]

[0087] In the formula: is the convolutional output of the i-th scale, representing the time series feature response at scale k i below.

[0088] Combining the outputs of convolutional kernels of different scales, the convolutional output of each scale is merged by weighted average:

[0089]

[0090] In the formula: where, ω i is the weight of each convolutional kernel, assigning different weights to the convolutional results of each scale to strengthen the features of a specific scale.

[0091] Combined with the adaptive convolution mechanism, it is dynamically adjusted according to the data features. To solve the problem that important information may be lost in long sequences of highly non-linear and complex patterns, bidirectional LSTM is applied to enable the model to learn from both the past and future information of the sequence, thereby capturing more detailed time-dependent relationships.

[0092] Introduce residual connections. By adding skip connections, information can be directly passed through certain layers, alleviating the problem of vanishing gradients.

[0093] Finally, adopt the L2 regularization method based on weight decay. By penalizing the magnitude of the weights, the complexity of the model is restricted to prevent overfitting in the case of a small amount of training data or large data noise.

[0094] Step 3 is specifically implemented as follows:

[0095] The present invention proposes a method for initializing the gray wolf population based on Tent-Logistic-Bernoulli multiple chaotic mappings. By utilizing the unpredictability and non-ergodicity of multiple chaotic sequences, a population with diversity and coverage is generated, enabling a more balanced exploration of the gray wolf population in the solution space, avoiding premature convergence of the algorithm, and enhancing the global search ability.

[0096]

[0097]

[0098]

[0099]

[0100] Where: are the initial populations generated by the Tent chaotic mapping, Logistic chaotic mapping, and Bernoulli chaotic mapping respectively, is the finally generated initial gray wolf population, and r and u are chaotic control parameters respectively.

[0101] In the original algorithm, the convergence factor a linearly decreases from 2 to 0 with the number of iterations. However, the algorithm is not linear during the continuous convergence process, that is, the linearly decreasing convergence factor a cannot fully reflect the actual optimization search process. Therefore, the present invention proposes an adaptive adjustment mechanism for the convergence factor based on the sine law to coordinate the global exploration and local development capabilities.

[0102]

[0103] Where: a initial and a final are the initial value and the final value of the convergence factor a, k is the current number of iterations, k max is the maximum number of iterations, and n is the decreasing exponent, where 0 < n ≤ 1.

[0104] The image of the improved convergence factor a is a curve that changes based on the sine law. It decreases slowly in the initial stage of iteration, enabling the convergence factor a to remain at a relatively large value for a long time to improve the search efficiency; it decreases rapidly in the later stage of iteration, enabling a to remain at a relatively small value for a long time to improve the search accuracy. Therefore, the global search and local search capabilities of the algorithm are balanced.

[0105] Aiming at the problems that in multi-peak or complex multi-objective optimization problems, the population is easily attracted by local optima and leads to premature convergence, and the concentration of the positions of the wolf packs during the iteration process causes insufficient diversity of solutions, an improved method based on a dynamic Gaussian mutation operator to enhance the search efficiency of the Pareto front is proposed. By adaptively adjusting the mutation intensity at different search stages, the global exploration and local development capabilities are effectively balanced.

[0106] Dynamic control formula for mutation probability based on iteration number attenuation:

[0107]

[0108] In the formula: P mutation (k) is the mutation probability of the k-th iteration; P start is the initial mutation probability; α is the parameter for controlling the descent speed.

[0109] Dynamic mutation operator based on Gaussian mutation:

[0110] X mutation = X current + σN(0,1) (34)

[0111] In the formula: X mutation is the position of the mutated gray wolf; X current is the position of the gray wolf before mutation; σ is the standard deviation that changes dynamically with iteration.

[0112] The beneficial effects of the present invention are as follows: First, a mathematical model of the integrated water-wind-solar-storage complementary system is established, and a multi-objective optimization framework for cascaded water-wind-solar-storage considering both power generation benefits and peak shaving effects is constructed. Subsequently, a day-ahead prediction framework for water-wind-solar using an adaptive multi-time scale Improved LSTNet is proposed. By using multi-scale convolutional kernels to design, features at different scales of time series are extracted, enhancing the model's ability to capture multi-level time series patterns such as different periodicities, trends, and randomness. At the same time, combined with an adaptive convolution mechanism, dynamic adjustment is performed according to data features. The predicted values of water-wind-solar are input into the multi-objective optimization framework. Finally, an improved multi-objective hybrid particle swarm gray wolf optimization algorithm is proposed to solve the established model. The gray wolf population is initialized using Tent-Logistic-Bernoulli multiple chaotic mappings. The global exploration and local development capabilities are coordinated through a convergence factor adaptive adjustment mechanism, and a dynamic Gaussian mutation operator is introduced to enhance the Pareto front search efficiency. Finally, a multi-objective day-ahead scheduling scheme for the complementary system is obtained. The proposed method achieves a good balance between peak shaving effects and power generation benefits, effectively reducing the peak-valley difference of the power grid and improving the stability of the power system. Description of the Drawings

[0113] Figure 1 is a schematic diagram of the cascaded water-wind-solar-storage complementary system;

[0114] Figure 2 is the power grid load diagram connected to the cascaded water-wind-solar-storage system;

[0115] Figure 3 is the day-ahead prediction data diagram of wind and solar;

[0116] Figure 4 It is the flow chart of IMOPSO - GWO solution;

[0117] Figure 5 It is the output process diagram of different cascade hydropower stations;

[0118] Figure 6 It is the power receiving process diagram of the power grid connected to the cascade water - wind - light - storage system;

[0119] Figure 7 It is the scheduling result diagram of the energy storage device;

[0120] Figure 8 It is the multi - objective optimization result diagram of IMOPSO - GWO. Specific implementation mode

[0121] The present invention will be described in detail below in conjunction with the accompanying drawings and specific implementation modes.

[0122] A multi - objective day - ahead peak - shaving scheduling method for a cascade water - wind - light - storage complementary system based on an improved hybrid particle swarm gray wolf optimization algorithm of the present invention is specifically implemented according to the following steps:

[0123] Step 1 is specifically implemented as follows:

[0124] Establish a peak - shaving objective function with the minimum average value of the absolute value of the surplus load deviation as the goal:

[0125]

[0126] In the formula: F1 is the sum of the average values of the absolute values of all grid surplus load deviations; G is the number of power grids; T is the total number of scheduling periods; L is the number of UHV DC tie lines; D g,t and D' g,t are respectively the system load and the surplus load of the g - th power grid at time t; P l,g,t is the power transmitted from the cascade water - wind - light - storage integrated complementary system to the g - th power grid through tie line l at time t; is the demand response power of the g - th power grid at time t.

[0127] Establish a power generation objective function with the maximum power generation of the cascade hydropower station as the goal:

[0128]

[0129] In the formula: F2 is the total power generation of the cascade hydropower; I is the number of cascade hydropower stations; is the output of the n - th unit of hydropower station i at time t; N i is the number of units of hydropower station i.

[0130] Mathematical models are established for the UHV DC tie line constraint, cascade hydropower constraint, energy storage constraint, and demand response constraint in the multi-objective optimization model of the integrated water-wind-solar-storage complementary system, respectively.

[0131] The UHV DC tie line constraint includes the upper and lower power limits and the amplitude change constraint. The cascade hydropower constraint includes the water volume balance constraint, reservoir water level limit constraint, outflow discharge constraint, power generation limit constraint of the power station, water level-storage capacity relationship, tail water level-outflow discharge relationship, unit generating head constraint, head loss function, unit output characteristic relationship, unit output constraint, and unit output ramp constraint. The energy storage constraint includes the energy storage power constraint, capacity constraint, and charge-discharge frequency constraint.

[0132] Step 2 is specifically implemented as follows:

[0133] Collect relevant historical data and meteorological data for the prediction of cascade water-wind-solar, including: historical hydrological data of the basin area, power generation of photovoltaic power plants and wind farms, rainfall, light intensity, wind speed, temperature, etc., and organize the above data into the same time window. Subsequently, outlier removal and missing value filling are performed. In the present invention, the Z-score method is used to perform outlier removal processing on the collected data, and when the relative difference is greater than 3, the data is removed as an outlier.

[0134]

[0135] In the formula: X is the sample data, μ is the sample mean, and σ is the standard deviation.

[0136] For blank values, inverse distance weighted interpolation is used and combined with the reasonable guidance range of meteorological data and expert experience screening for filling. Since the above time series data has significant seasonal fluctuations, masking the underlying trends or periodic changes. Therefore, through STL seasonal decomposition, the seasonal fluctuations are separated from the original data, making the model more focused on the data trends and residuals, thereby improving the prediction accuracy.

[0137] The core of the STL method lies in decomposing the time series Y t into three parts. The trend component is estimated by smoothing the original data through Loess regression; the seasonal component is obtained by Loess smoothing after removing the trend component; the residual component is the result of subtracting the trend and seasonal components from the original data.

[0138] Y t = T t + S t + R t (38)

[0139] T t = Loess(Y t ) (39)

[0140] S t = Loess(Y t - T t ) (40)

[0141] R t = Y t - T t - S t (41)

[0142] Where: Y t is the original time series data; T t is the trend component; S t is the seasonal component; R t is the residual component.

[0143] To make the training of the Improved LSTNet model more stable, the Min - Max normalization method is applied to normalize the input data.

[0144] Using the multi - scale convolutional kernel design to extract features of different scales of the time series, enhancing the model's ability to capture multi - level time series patterns such as different periodicities, trends, and randomness.

[0145] Combined with the adaptive convolution mechanism, it is dynamically adjusted according to the data features. Different from the traditional fixed convolutional kernel, the adaptive convolution operation can automatically adjust the size, weight, or stride of the convolutional kernel according to different characteristics of the data, improving the model's performance in different data environments. Through reinforcement learning, the adaptive convolutional kernel size k adaptive , the adaptive convolution weight K adaptive , and the adaptive stride s adaptive are generated respectively.

[0146] k adaptive = f size (X) (42)

[0147] K adaptive = f weight (X) (43)

[0148] s adaptive = f stride (X) (44)

[0149] Where: X = {x1, x2,... x T} is the input sequence, where each is the input vector at time t, d is the input feature dimension, and T is the sequence length; f size (X), f weight (X), f stride (X) are the reinforcement learning modules respectively.

[0150] Based on the above three adaptive mechanisms, the convolution operation is as follows:

[0151]

[0152] In the formula: y t represents the convolution output.

[0153] To address the problem that highly non-linear and complex pattern long sequences may lose important information, bidirectional LSTM is applied to enable the model to learn from both past and future information of the sequence simultaneously, thereby capturing more detailed temporal dependencies. Residual connections (ResNet) are introduced, and by adding skip connections, information can be directly transmitted bypassing certain layers, alleviating the problem of gradient disappearance. The L2 regularization method based on weight decay is adopted. By penalizing the magnitude of the weights, the complexity of the model is restricted to prevent overfitting in cases where the training data volume is small or the data noise is large.

[0154] After using the adaptive multi-time scale Improved LSTNet neural network to perform day-ahead prediction on the water, wind, and light, the predicted data is input into the cascade water, wind, light, and storage multi-objective optimization framework.

[0155] Step 3 is specifically implemented as follows:

[0156] The present invention proposes a method for initializing the gray wolf population based on Tent-Logistic-Bernoulli multiple chaotic mappings. Utilizing the unpredictability and non-ergodicity of multiple chaotic sequences, a population with diversity and coverage is generated, enabling the gray wolf population to explore more evenly in the solution space, avoiding premature convergence of the algorithm, and enhancing the global search ability.

[0157]

[0158]

[0159]

[0160]

[0161] In the formula: are the initial populations generated by the Tent chaotic mapping, Logistic chaotic mapping, and Bernoulli chaotic mapping respectively, is the finally generated initial gray wolf population, and r and u are chaotic control parameters respectively.

[0162] In the original algorithm, the convergence factor a linearly decreases from 2 to 0 with the number of iterations. However, the algorithm is not linear during the continuous convergence process, that is, the linearly decreasing convergence factor a cannot fully reflect the actual optimization search process. Therefore, the present invention proposes an adaptive adjustment mechanism for the convergence factor based on the sine law to coordinate the global exploration and local development capabilities.

[0163] A = 2ar1 - a (50)

[0164] C = 2r2 (51)

[0165]

[0166] Where: A and C are coefficient vectors; the magnitudes of r1 and r2 are random numbers between [0, 1]; a initial and a final are the initial value and the final value of the convergence factor a; k is the current iteration number; k max is the maximum iteration number; n is the decreasing exponent, 0 < n ≤ 1.

[0167] The image of the improved convergence factor a is a curve that changes based on the sine law. It decreases slowly in the initial stage of iteration, so that the convergence factor a remains a relatively large value for a long time to improve the search efficiency; it decreases rapidly in the later stage of iteration, so that a remains a relatively small value for a long time to improve the search accuracy. Therefore, it balances the global search and local search capabilities of the algorithm.

[0168] The wolf pack continuously approaches the prey under the leadership of the α, β, and δ wolves. During this process, their positions are always changing dynamically until the hunting is successful. This process is represented by the following formula:

[0169]

[0170]

[0171]

[0172] ω (t) =(ω ini -ω end )(k max -k) / k max +ω end (56)

[0173] Where: D α 、D β 、D δ are the distances between other individuals in the wolf pack and α, β, and δ respectively; X α 、X β 、X δThe current positions of wolves α, β, and δ respectively; X1, X2, and X3 are the positions that other individuals in the population need to adjust under the influence of wolves α, β, and δ respectively; c1 and c2 represent the individual learning factor and the global learning factor respectively; ω represents the inertia weight, and a linear decreasing weight strategy is adopted, starting from the initial value ω ini Linearly decrease to ω end .

[0174] After each hunting of the wolf pack, the non-dominated optimal solution set is updated. An external population Archive mechanism is introduced to store the non-dominated optimal solutions. A leader selection strategy is adopted to select the leader during the hunting process from the external population Archive. A grid mechanism is applied to measure the crowding degree of the non-dominated solution set.

[0175] Aiming at the problems that in multi-modal or complex multi-objective optimization problems, the population is easily attracted by local optima and leads to premature convergence, and the lack of diversity of solutions caused by the centralization of the wolf pack positions during the iteration process, an improved method based on a dynamic Gaussian mutation operator to enhance the Pareto front search efficiency is proposed. By adaptively adjusting the mutation intensity at different search stages, the global exploration and local exploitation capabilities are effectively balanced.

[0176] Dynamic control formula for mutation probability based on the decay of the number of iterations:

[0177]

[0178] In the formula: P mutation (k) is the mutation probability of the k-th iteration; P start is the initial mutation probability; α is the parameter controlling the descent speed.

[0179] Dynamic mutation operator based on Gaussian mutation:

[0180] X mutation = X current + σN(0,1) (58)

[0181] In the formula: X mutation is the position of the mutated gray wolf; X current is the position of the gray wolf before mutation; σ is the standard deviation that changes dynamically with the iteration.

[0182] Embodiment

[0183] Taking the complementary system constructed with reference to 15 units of 4 hydropower stations, a wind farm, and a photovoltaic power station in a cascade of a certain basin in Southwest China as the research object, the simplified cascade hydraulic relationship and network topology are as Figure 1 shown. The four hydropower stations are respectively connected to a section to transmit electric energy to the power grid of a region. The loads of the four power grids are as Figure 2 shown.

[0184] First, a mathematical model of the integrated water-wind-solar-storage complementary system is established, and a multi-objective optimization framework for cascaded water-wind-solar-storage considering power generation benefits and peak shaving effects is constructed. Subsequently, an adaptive multi-time scale Improved LSTNet day-ahead prediction framework is used to predict water, wind, and solar power. The output of wind and solar power is as Figure 3 shown.

[0185] The day-ahead prediction values of water, wind, and solar power are input into the multi-objective optimization model, and an improved multi-objective hybrid particle swarm grey wolf optimization algorithm is used to solve the established model. The solution process is as Figure 4 shown.

[0186] The output process of the cascaded hydropower stations is as Figure 5 shown. From the output process, after considering the constraints of unit start-stop and fluctuation duration, the output of each power station is relatively stable, without frequent fluctuations, meeting the actual operation requirements of the power station. In addition, the cascaded hydropower stations have high adaptability to the response of grid load demand and can effectively meet the requirements of operation scheduling for stability and controllability.

[0187] Table 1 and Figure 6 show in detail the output distribution, hydropower consumption, and peak shaving response results of the cascaded hydropower stations in the grid during the peak shaving process.

[0188] Table 1 shows the received power and peak shaving effects of each grid. Among them, the received power of Grid IV is the largest, reaching 19077 MW·h, and the received power of Grid III is the smallest, only 8199 MW·h. After receiving the power transmission from the hydropower station, the remaining load of each grid tends to be stable, and the peak-valley difference is effectively adjusted. The analysis results show that Grid I has the best peak shaving effect, with the peak-valley difference reduced by 77.15% and the average value of the absolute value of the anomaly reduced by 90.31%. The peak shaving effect of Grid IV is the second best, with the average value of the absolute value of the anomaly reduced by 90.05%. Comprehensive analysis of the grid load characteristics shows that Grid I has a relatively small peak-valley difference of the load, and Hydropower Station I can effectively suppress the load fluctuation. Although Grid IV has the largest peak-valley difference of the load, due to its five units and large installed capacity, it can also effectively smooth the remaining load. In contrast, Grid II and Grid III have relatively large peak-valley differences of the load, and the corresponding installed capacity of the cascaded hydropower stations is small, resulting in their transmission power being unable to accurately track the load changes, thus facing greater peak shaving pressure.

[0189] Table 1 Peak shaving effect indicators

[0190]

[0191] The scheduling results of the energy storage devices of the photovoltaic power station and the wind farm are as Figure 7As shown in the figure, IMOPSO-GWO can obtain a given number of solution sets and obtain representative scheduling schemes. Given the number of solutions as 30, that is, it is expected to obtain 30 sets of scheduling schemes. Figure 8 It is the multi-objective optimization result obtained by IMOPSO-GWO. The power grid peak shaving target and the total power generation target of cascade hydropower stations restrict each other and there are significant contradictions. On the Pareto front, as the power grid peak shaving effect improves, the total power generation of cascade hydropower stations decreases. Moreover, after applying the grid mechanism and the leader selection mechanism, the Pareto solution set is also more evenly distributed, and each solution has high representativeness, providing a reliable decision-making basis for engineering practical applications.

[0192] It can be understood that the present invention is described through some embodiments. Those skilled in the art know that without departing from the spirit and scope of the present invention, various changes or equivalent replacements can be made to these features and embodiments. In addition, under the teaching of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the scope protected by the present invention.

Claims

1. A multi-objective day-ahead peak-shaving scheduling method for a cascaded hydro-wind-solar-storage complementary system based on an improved hybrid particle swarm grey wolf optimization algorithm, characterized in that: Follow the steps below to implement it: Step 1: Establish a mathematical model of the integrated complementary system of water, wind, solar and storage, comprehensively consider the constraints of UHV DC interconnection lines, cascade hydropower constraints, energy storage constraints, and demand response constraints, and construct a multi-objective optimization framework for cascade water, wind, solar and storage that considers power generation benefits and peak-shaving effects. Step 2: Propose an adaptive multi-time scale Improved LSTNet water, wind and solar day-ahead forecasting framework. Use multi-scale convolution kernel design to extract features of different scales of time series, and enhance the model's ability to capture multi-level time series patterns such as different periodicity, trend, and randomness. At the same time, combine the adaptive convolution mechanism to make dynamic adjustments based on data characteristics. Step 3. An improved multi-objective hybrid particle swarm grey wolf optimization algorithm (IMOPSO-GWO) is proposed to solve the constructed model. The grey wolf population is initialized using the Tent-Logistic-Bernoulli multiple chaotic mapping. The global exploration and local development capabilities are coordinated through the adaptive adjustment mechanism of the convergence factor. The dynamic Gaussian mutation operator is introduced to enhance the Pareto frontier search efficiency. Finally, a multi-objective day-ahead scheduling scheme for the complementary system is obtained.

2. The multi-objective optimization framework for cascade hydropower, wind power, solar power and storage considering power generation benefits and peak load regulation effects according to claim 1 is characterized in that: The step 1 is specifically implemented as follows: The multi-objective optimization model of the integrated complementary system of water, wind, solar and storage includes the peak regulation effect and power generation objective functions. Objective function 1: the average value of the absolute value of the residual load deviation is the minimum Where: F1 is the sum of the average values ​​of the absolute values ​​of the residual load deviations of all power grids; G is the number of power grids; T is the total number of dispatching periods; L is the number of UHV DC interconnection lines; D g,t and D' g,t are the system load and residual load of the g power grid in time period t respectively; P l,g,t The power transmitted from the interconnection line l to the grid g in time period t by the cascade hydro-wind-solar-storage integrated complementary system; is the demand response power of grid g in period t. Objective function 2: Maximum cascade hydropower generation Where: F2 is the total power generation of cascade hydropower; I is the number of cascade hydropower stations; is the output of the nth unit of hydropower station i in time period t; N i is the number of units in hydropower station i. The multi-objective optimization model of the integrated complementary system of water, wind, solar and storage includes the following types of constraints: UHV DC interconnection line constraints, cascade hydropower constraints, energy storage constraints, and demand response constraints. UHV DC interconnection line constraints include power upper and lower limit constraints and amplitude constraints. Cascade hydropower constraints include water balance constraints, reservoir water level constraints, outflow constraints, power station output constraints, water level-reservoir capacity relationship, tailwater level-discharge flow relationship, unit generating head constraints, head loss function, unit output characteristic relationship, unit output constraints, and unit output ramp constraints. Energy storage constraints include energy storage power constraints, capacity constraints, and charge and discharge frequency constraints.

3. The water, wind and solar day-ahead prediction framework based on adaptive multi-time scale Improved LSTNet according to claim 1 is characterized in that: The step 2 is specifically implemented as follows: (1) Data preprocessing Collect relevant historical data and meteorological data for cascade water, wind and light forecasts, including: historical hydrological data of the basin area, power generation of photovoltaic power stations and wind farms, rainfall, light intensity, wind speed, temperature, etc., and organize the above data into the same time window. Subsequently, outliers are removed and missing values ​​are filled. The present invention uses the Z-score method to perform abnormal elimination processing on the collected data and sets that when the relative difference is greater than 3, the data is removed as an outlier. For blank values, inverse distance weighted interpolation is used and filled in combination with the reasonable guidance range of meteorological data and expert experience screening. Since the above time series data has significant seasonal fluctuations, the potential trends or periodic changes therein are concealed. Therefore, through STL seasonal decomposition (Seasonal-Trend Decomposition using LOESS), seasonal fluctuations are separated from the original data, so that the model focuses more on data trends and residuals, thereby improving the prediction accuracy. In addition, in order to make the training of the ImprovedLSTNet model more stable, the Min-Max normalization method is used to normalize the input data. (2) Construct and train the adaptive multi-time scale Improved LSTNet prediction network The multi-scale convolution kernel design is used to extract features of different scales of time series, and enhance the model's ability to capture multi-level time series patterns such as different periodicity, trend, and randomness. At the same time, the adaptive convolution mechanism is combined to make dynamic adjustments according to data characteristics. In order to solve the problem that long sequences with highly nonlinear and complex patterns may lose important information, the bidirectional LSTM is applied to enable the model to learn from the past and future information of the sequence at the same time, thereby capturing more detailed time dependencies. Residual connections are introduced, and skip connections are added to enable information to bypass certain layers and pass directly, alleviating the problem of gradient disappearance. Finally, the L2 regularization method based on weight decay is used to limit the complexity of the model by penalizing the size of the weights, preventing overfitting when the amount of training data is small or the data noise is large.

4. The IMOPSO-GWO algorithm according to claim 1, characterized in that: The step 3 is specifically implemented as follows: (1) When the solution space of the optimization problem is large, the conventional random initialization of the population will cause the population to be concentrated in certain local areas of the solution space. Therefore, the present invention proposes a method for initializing the gray wolf population based on the Tent-Logistic-Bernoulli multiple chaotic mapping. By utilizing the unpredictability and non-traversability of multiple chaotic sequences, a population with diversity and coverage is generated, so that the exploration of the gray wolf population in the solution space is more balanced, the algorithm is prevented from converging prematurely, and the global search capability is enhanced. Where: are the initial populations generated by Tent chaotic mapping, Logistic chaotic mapping, and Bernoulli chaotic mapping, respectively. is the final generated initial population of gray wolves, r and u are the chaos control parameters respectively. (2) In the original algorithm, the convergence factor a decreases linearly from 2 to 0 with the number of iterations, but the algorithm is not linear in the process of continuous convergence, that is, the linearly decreasing convergence factor a cannot fully reflect the actual optimization search process. Therefore, the present invention proposes a convergence factor adaptive adjustment mechanism based on sinusoidal changes to coordinate global exploration and local development capabilities. Where: a initial and a final is the initial value and final value of the convergence factor a, k is the current number of iterations, k max is the maximum number of iterations, n is the decreasing index, 0 <n≤1。 The improved convergence factor a is a curve based on the sine law. It decreases slowly in the early stage of iteration, so that the convergence factor a maintains a larger value for a longer time to improve the search efficiency; it decreases quickly in the later stage of iteration, so that a maintains a smaller value for a longer time to improve the search accuracy. Therefore, the global search and local search capabilities of the algorithm are balanced. (3) In order to address the problem that in multi-peak or complex multi-objective optimization problems, the population is easily attracted to the local optimum, resulting in premature convergence, and the wolf pack's position is centralized during the iteration process, resulting in insufficient solution diversity, an improved method based on a dynamic Gaussian mutation operator to enhance the Pareto frontier search efficiency is proposed. By adaptively adjusting the mutation intensity at different search stages, the global exploration and local development capabilities can be effectively balanced. Dynamic control formula of mutation probability based on attenuation of iteration number: Where: P mutation (k) is the mutation probability of the kth iteration; P start is the initial mutation probability; α is the parameter that controls the decline speed. Dynamic mutation operator based on Gaussian mutation: X mutation =X current +σN(0,1) (9) Where: X mutation is the position of the gray wolf after mutation; X current is the position of the gray wolf before mutation; σ is the standard deviation that changes dynamically with iteration.

Images